anthropic CEO Dario Amodi made one Necessary shock In April, for the need to understand how AI models think.

It comes at an important time. Anthropic battle In the Global AI rankings, it is important to note what sets it apart from other top AI laboratories. Since its establishment in 2021, when seven Openi Employee Broke On concerns about AI safety, Anthropic has created the AI ​​model that follows a set of human-valuable principles, a system they say Constitutional AIThese principles ensure that models are “are”Auxiliary, honest and harmless“And usually act in the best interests of society. At the same time, anthropic is research branch Deep diving To understand how its models think about the world, and Why They produce auxiliary (and sometimes harmful) answer.

Anthropic flagship model, Cloud 3.7 sonnet, Domination When launching in February, benchmark coding, proving that AI models can excel in both performance and security. And recently the release of Cloud 4.0 Opus and Sonnet again puts Cloud again Above the coding benchmarkHowever, in today’s rapid and hyper-replicated AI market, GOOGLE’s Gemini 2.5 Pro and Open AI’s O3 have their impressive performances for coding for coding, while they are while they are Already dominated Cloud in mathematics, creative writing and overall logic in many languages.

If Amodi’s views are no signs, then anthropic is planning for the future of AI and its implications in important areas such as drug, psychology and law, where models safety and human values ​​are essential. And it shows: Anthropic is a major AI lab that focuses strictly on developing “interpretable” AIs, which are models that are models, to some extent for certainty, for certainty, what the model is thinking and how it comes to a particular conclusion.

Amazon and Google Already has invested billions of dollars in anthropic, even they manufacture their own AI models, so perhaps the competitive advantage of anthropic is still debutant. Explanative model, as an anthropic suggestions, can significantly reduce long -term operating costs associated with debugging, auditing and rising risks in complex AI deployment.

Sayash KapoorAn AI security researcher, suggesting that the lecturer is valuable, is one of the many devices for the management of AI risk. In his view, “lecturer is neither necessary nor enough” to safely behave the model safely “-This filter is most important when it is added with verifier and human-focused design. It sees the more expander visual lecturer as part of a large ecosystem of control strategies, especially in real -world AI -finance where models are components in comprehensive decision -making systems.

Interpretable AI requirement

Until some time ago, many people thought Bangle Adopt extraordinary markets. While these models are already Push the front of human knowledgeTheir comprehensive use is only responsible for how good they are in solving a wide range of practical problems, which requires creative problem-solution or detailed analysis. As the model is placed for work on rapidly important problems, it is important that they produce an accurate answer.

Amodei is afraid that when AI responds to a hint, “We don’t know … why it chooses some words on others, or why ever it makes a mistake despite being accurate.” Such errors – hallucinations of misinformation, or reactions that do not align with human values ​​- the AI ​​model will catch back from reaching their full capacity. In fact, we have seen many examples of AI Sad dream And unethical behavior,

For Amodei, the best way to solve these problems is to understand how AI thinks: “Our disability in understanding the internal mechanism of the model means that we cannot make such predictions meaningfully. [harmful] Behavior, and so struggle to rule them … If it was possible to look inside the model instead, we can all be able to systematically block the gelbreak, and also tell what the model has dangerous knowledge. ,

Amodei also sees the ambiguity of current models as a barrier to deploy the AI ​​model in high-day financial or safety-mating settings, as we cannot completely determine the limit on their behavior, and a small number of mistakes can be very harmful. ” In decision making that directly affects humans, such as medical diagnosis or mortgage assessment, legal Regulations AI is required to explain its decisions.

Imagine a financial institution using a large language model (LLM) to detect fraud – lecturer may mean to interpret the loan application refused to the customer required by the law. Or optimization of a manufacturing firm supply chains – to understand why an AI suggests that a particular supplier can unlock the ability and prevent unexpected bottlenecks.

Because of this, Amodei explains, “Anthropic interpretation is doubling, and we have a goal for ‘lecturers the most model problems’ by 2027.

By that end, Anthropic recently participated in $ 50 million Investment In GoodfireAI Research Lab AI is progressing success on “brain scan”. Their model inspection platform, amber, is an unknowable device that identifies the concepts learned within the model and allows users to manipulate them. recently DemoThe company showed how Amber an image can identify personal visual concepts within AI and then let users go paint These concepts on a canvas to generate new images that follow the user’s design.

The investment of anthropic in amber indicates on the fact that it is quite difficult to develop an explanatory model that there is no manpower to interpret it in anthropic. Creative explanatory models require new toolchen and skilled developers to create them

Extensive reference: perspective of an AI researcher

To break the perspective of Amodei and add very important references, Venturebeat interviewed Kapoor an AI security researcher in Princeton. Kapoor co-written the book AI snake oilA significant examination of exaggerated claims around the capabilities of the major AI model. He is also a co-writer of “Aye as a general technique“In which he advocates assuming AI as a standard, transformative tool such as internet or electricity, and promotes a realistic perspective on its integration in everyday systems.

Kapoor does not dispute that interpretation qualification is valuable. However, he is suspected of considering it as the central column of AI alignment. “It’s not a silver bullet,” Kapoor told Venturebeat. He said that many of the most effective safety techniques, such as filtering, do not need to open the model, he said.

He also warns what researchers have called the “decline of disqualification” – the idea that if we do not fully understand the internal of a system, we cannot use or regulate it responsibly. In practice, complete transparency is not how most technologies are evaluated. What matters whether a system performs firmly in real circumstances.

This is not the first time Amodei has warned of AI’s risks to overcome our understanding. In your October 2024 Post“Machines of Laving Grace,” they sketched a vision of the fast enabled model, which can double the meaningful real -world action (and perhaps double our lifetime).

According to Kapoor, there is a significant difference between a model here Capacity And its PowerModel abilities are undoubtedly growing rapidly, and they may soon develop enough intelligence to find solutions to many complex problems challenging humanity. But a model is only as powerful as the interface we provide it to interact with the real world, where and how the models are involved.

Amodei has argued that America should maintain a lead in AI development, in part, export control This limits access to powerful models. The idea is that the totalitarian governments can use the Frontier AI system irresponsibly -or can seize geo -political and economic growth that come with them first deployed.

For Kapoor, “Even the greatest proposals of export control agree that it will give us in a year or two.” He thinks that we should treat AI as “General technology“Like lightning or internet. Remaining, it took both technologies to feel fully in society. Kapoor feels that it is the same for AI: The best way to maintain geo -political edge is to focus on” long games “to change industries to effectively use industries.

Other people criticized Amodei

Kapoor is not just a criticism of Amodi. Last week in Paris in Wivatech, Jenson Huang, NVDia CEO, Announced her disagreement With the ideas of Amodei. Huang questioned whether the right to develop AI should be limited to some powerful institutions like Anthropic. He said: “If you want things to be done safely and responsible, then you do it in the open … Do not do it in a dark room and tell me that it is safe.”

In response, anthropic Stated: “Dario has never claimed that ‘only anthropic’ can create a safe and powerful AI. As the public records show, Dario has advocated a national transparency standard for AI developers (including anthropic), so public and policy maker are aware of models’ abilities and risks and can prepare it accordingly.”

It is also worth noting that anthropic is not alone in search of its interpretation: Google’s Deepmind Enterprise team has also created under the leadership of Neil Nanda. Serious contribution For lecturer research.

Ultimately, the top AI labs and researchers are providing strong evidence that the lecturer in the competitive AI market can be a major difference. Enterprises that prefer lecturers can gain a significant competitive edge by creating a more reliable, obedient and optimal AI system.